Handle ConstantFP vector splats in transConstant (#3705)
llvm/llvm-project@d19e954b83 made `-use-constant-fp-for-fixed-length-splat` default to true.
This causes `ConstantFP` to represent fixed-length vector splats natively instead of using `ConstantDataVector`.
The translator only handled scalar `ConstantFP`, so vector-typed `ConstantFP` splats were silently mistranslated.
e.g., this LLVM IR call argument:
`call <8 x i8> @convert(<8 x half> <half 0xH3C00, half 0xH3C00, ...>)`
produced a scalar `OpConstant` with a vector type:
`%53 = OpConstant %v8half 15360 ; invalid: scalar op, vector type`
instead of the correct `OpConstantComposite`:
```
%53 = OpConstant %half 15360
%60 = OpConstantComposite %v8half %53 %53 %53 %53 %53 %53 %53 %53
```
Build an `OpConstantComposite` from repeated scalar references instead.
Existing tests that cover this:
- extensions/EXT/SPV_EXT_float8/conversions_scalar_vector.ll
- extensions/INTEL/SPV_INTEL_float4/conversions_packed.ll
- extensions/INTEL/SPV_INTEL_float4/conversions_scalar_vector.ll
- extensions/INTEL/SPV_INTEL_int4/conversions_packed.ll
AI-assisted: Claude Sonnet 4.6 (commercial SaaS)
Original commit:
https://github.com/KhronosGroup/SPIRV-LLVM-Translator/commit/d4d65d5ea9449ae
